Obstacle Avoidance For Mobile Robot Based On Stereo Vision

Stereo vision is a technology which uses two cameras to calculate the distance between camera and object.Compared with monocular camera,the stereo camera can provide the actual size of object.Thus,it has been widely applied in the field of robot navigation,3D reconstruction and target recognition.This thesis mainly focuses on the application of stereo vision in the field of robotic.First of all,GPU accelerated depth calculation is presented.In this chapter,the mathematic model of stereo camera is firstly introduced to lay the mathematic foundation of the remaining part of this thesis.After that,disparity calculation,including stereo calibration,GPU accelerated stereo rectification and stereo matching are discussed.Especially,GPU accelerated SGM algorithm is designed to meet the real-time requirement needed for AGV/UAV obstacle avoidance,Secondly,obstacle detection based on depth and RGB information is introduced.Inspired by the idea of salient object detection,a multi-feature-fused saliency detection method is raised in this thesis.This method fuses the depth prior into the manifold ranking algorithm,which can make the algorithm more accurate in the natural scene,especially on the cruising path of UAV/UGV.Thirdly,real-time obstacle detection and collision avoidance for UAV using depth and optical flow is discussed.Obstacle segmentation and avoidance algorithm based on stereo vision method is introduced.The self-adapted threshold method is designed to eliminate interference of ground for segmentation.In addition,an algorithm which merges the depth image and optical flow method is introduced to get the relative velocity of obstacle towards UAV.Based on this,a simple and fast avoidance strategy is designed.At last,the method on the navigation of unmanned ground vehicle(UGV)is studied.In this chapter,Convolutional Neural Network(CNN),which merges the depth and color information,is used to realize the autonomous cruise as well as obstacle avoidance of UGV in the known environment.Through large amount of experiments,it can be concluded that the obstacle recognition and autonomous navigation algorithm designed in this thesis achieves a higher accuracy and has a certain theoretical innovation and engineering application value.